Dynamical-bridge fidelity as a pre-release standard for hybrid seasonal tropical-cyclone forecasts

Hybrid machine-learning–physics models make seasonal tropical-cyclone ensembles computationally tractable, but regional risk products can be skillful for the wrong reason. Basin metrics do not reveal whether a forecast preserves the source of predictability, translates it through the correct circulation, or benefits from compensation among genesis and track errors. This Perspective proposes a pre-release audit that traces boundary memory through atmospheric mediation to the declared hazard. The verification unit—initialization, lead, season, region, tracker, and hazard definition—is fixed before scoring. Each link is evaluated for source fidelity, conditional translation, and counterfactual necessity on a common hindcast calendar using transparent baselines, targeted ablations, and negative controls. Western North Pacific evidence shows why the El Niño–Southern Oscillation, the Pacific and Atlantic meridional modes, and Indian Ocean variability are candidate sources rather than stand-alone predictors: value depends on realistic monsoon-trough, subtropical-ridge, shear, and steering responses. An observation-based 1979–2023 partition defines regional targets and exposes compensation within basin totals; it is not a model validation. Release should require added value across all three links, calibration for the communicated hazard, and stability under climate-state stress tests.